Nine megawatts into eight

The rack is the NVIDIA GB300 NVL72: 72 Blackwell Ultra accelerators and 36 Grace host processors, per NVIDIA's published configuration. It draws 140 kW nominal at the rack input, per OEM system specifications. A 9 MW requirement is 64 whole racks and 4,608 accelerators, drawing just under 9 MW. The 8 MW envelope, IT load at the rack input, holds 57 racks and 4,104 accelerators today. The requirement is short by 7 racks and 504 accelerators. It needs about 11% off at the rack input: your arithmetic, with no number of ours in it.

Our coefficient is measured at the accelerator die. Your constraint sits at the rack input. Between them is the share of rack draw the accelerators account for, and that share is yours to state. Skip that step and a model overstates the reduction. On this rack the accelerator dies are about 101 kW of a 140 kW rack, which is 72%.

Both readings of the measurement travel together on every figure. The conservative reading applies 21% to the accelerator share only: 21% of 72% is about 15% at the rack input, modelled. The stated basis treats the die figure as a floor on wall watts, because a watt not drawn at the die is not converted and not cooled. That gives 21% at the rack input, modelled.

On the conservative reading, the 64 racks sit inside 8 MW with about 395 kW of headroom. The envelope now holds 67 racks and 4,824 accelerators. That is 3 racks and 216 accelerators beyond the requirement. The margin over the 11% needed is about 4 percentage points, roughly fifteen times our measurement uncertainty at the rack. Plan against this reading.

On the stated basis the envelope holds 72 racks and 5,184 accelerators: 8 racks and 576 accelerators beyond the requirement. Headroom is about 922 kW and the margin about 10 percentage points. The deal fits under either reading; the difference is what is left over. Per rack the reduction returns about 21 kW to about 29 kW. The site's rack count rises about 18% to 26%, modelled.

The freed capacity is spent once. Expand the same client's deployment by 216 to 576 accelerators, let the racks to a second tenant, hold them as headroom, or bank the energy. The first two use the same racks, so count one. Take an illustrative GBP 3 an accelerator-hour at 85% utilisation, both your own figures. The spare accelerators then gross about GBP 4,800,000 a year on the conservative reading and about GBP 13,000,000 on the stated basis, modelled.

The racks that earn it are hardware you buy: about GBP 2,600,000 delivered per rack, or GBP 7,800,000 to GBP 20,800,000 of capital. The runtime is software with no capital cost; what it removes is the power constraint. The energy line is the smallest pool and needs your own measured mean draw. At an illustrative 95 kW per rack and 20 pence a kilowatt-hour it is about GBP 1,600,000 to GBP 2,200,000 a year, modelled. Those are the same watts as the headroom, so bank one use, not both.

Requirement size moves it first. On the same 8 MW envelope the boundary sits between 67 and 68 racks. At 67 the conservative margin is under half a percentage point, too thin to plan a tenancy against. A validation on your own fleet comes before any commitment there. From 68 racks only the stated basis closes the gap. From 73 racks neither does, and the answer is a larger envelope or fewer racks.

The plane of the envelope moves the answer most. If your 8 MW is total facility power, divide it by your power usage effectiveness first: the stated basis still fits at 1.10, the conservative reading needs a ratio below about 1.05, and at 1.15 neither fits, so confirm the plane before anything else. Rack draw basis comes next. At the 155 kW peak figure only the stated basis fits, and at the 192 kW busway figure neither reading does. A sustained reduction relieves a contractual demand figure, not a plant rating, so read your envelope off the contract or the switchgear schedule first.

ADAPT installs as a Kubernetes and Helm runtime above the NVML interface, driver-adjacent, one lightweight service per accelerator. It needs no application code change and adds no draw of its own. It runs under existing MIG and vGPU tenant isolation and never touches placement or scheduling. It reports through standard Prometheus and DCGM metrics. Every clock decision passes a layered safety governor and commits only on a successful NVML write. Control releases the clock if latency or throughput slips.

The mechanism is duty-cycle intelligence on bursty fleets, mostly idle-floor reduction, so a fleet at full load all year returns less. Seven days of read-only 1 Hz telemetry shows where yours sits before any trial. This scenario runs the Balanced preset, the mode the coefficient was measured in. The run held throughput, 0% change, and gave up to +22% tokens per watt, both derived: the same tokens for less energy.

If a deployment occupies a site by virtue of this reduction, removing the software puts the site back over its limit. Continued operation is then a standing condition of the configuration. The runtime fails open to full performance on a component fault. Here that returns the deployment to just under 9 MW against 8 MW, about 960 kW over. The mitigation lives at the site: headroom held back rather than let, a curtailment right, an interlock, or a commitment set below the modelled figure. The conservative headroom of about 395 kW covers about 41% of the overdraw. Settle it in the first conversation.